forget-rag

A forgetting layer for RAG systems — heat-based decay, tiered storage, and graceful eviction so your knowledge base doesn't get dumber over time.

Status: v0.1 (alpha). BM25 + heat ranking, L1/L2/L3 tiers, soft delete with restore-ready schema. Vector layer lands in v0.2.

Install

pip install forget-rag

# with the LangChain adapter
pip install "forget-rag[langchain]"

Quickstart

from forget_rag import ForgettingMemory

memory = ForgettingMemory(
    backend="sqlite",
    decay_halflife_days=30,
    tiers={"L1": 1000, "L2": 10000, "L3": "unlimited"},
)

memory.add("Some knowledge chunk", tags=["meeting", "2026-Q1"])
results = memory.search("query", limit=5)   # auto-promotes hot chunks
report = memory.health_check()              # what should be forgotten?

What's in the box

• Heat score — every chunk has a decay function based on access frequency + recency.
• L1 / L2 / L3 tiers — modeled after CPU cache; hot stays in vector
  • FTS, warm in FTS only, cold is archived but searchable.
• Soft delete — forget() flags rather than drops; v0.2 will expose restore() on top of the existing forgotten_at column.
• LangChain adapter — ForgettingRetriever extends BaseRetriever for drop-in chain use.

Companion CLI

The mem-broom companion package wraps this library with a typer-based CLI for ad-hoc memory hygiene. v0.1 ships it as a git install (the PyPI publish is deferred — see the project README for the install command).

Links

• Repo: https://github.com/zx22413/forget-rag
• Benchmark write-up: https://github.com/zx22413/forget-rag/blob/main/docs/benchmark.md
• Architecture: https://github.com/zx22413/forget-rag/blob/main/docs/architecture.md
• Issues: https://github.com/zx22413/forget-rag/issues

License

MIT.